Interrater reliability for Dirichlet, compositional, and Euclidean data
摘要
No metric has previously been developed to assess reliability across multiple Dirichlet, compositional, or Euclidean samples. Many research applications, such as latent Dirichlet allocation and other topic modeling techniques, would benefit from a formal method to assess reliability in order to serve purposes such as cross-validation. In this manuscript, a family of statistical approaches is presented to estimate category-wise and omnibus reliability for Dirichlet, compositional, or Euclidean data with K ≥ 2 dimensions. A general approach is presented for any data that may be conceived as representing positions in Euclidean space, including compositional data, after which a more tailored approach is offered to account for the idiosyncrasies of Dirichlet-distributed data. These approaches cleanly map onto existing reliability coefficients that are routinely used to assess univariate data.